UC Berkeley

This dissertation addresses methodological and substantive questions around the estimation of contextual effects, with a focus on the influence of community factors on alcohol use in South Africa. Alcohol use contributes to a multitude of health conditions and is a particularly important risk factor for South African men due to heavy consumption and the high risk of alcohol-related harms such as violence and road accidents in this population. In addition, HIV transmission and progression to AIDS may be accelerated by alcohol use. Research into the causes of dangerous alcohol use to date has focused on individual-level factors. This work addresses contextual traits that may relate to alcohol use, particularly community social factors. Specifically, I evaluate analytic methods for the measurement and aggregation of individual perceptions of social factors such as collective efficacy (chapter 2), test the associations of alcohol outlet density and village collective efficacy with young men’s drinking behaviors in rural South Africa (chapter 3), and assess the village-level effect of a randomized community mobilization trial intended to modify gender norms on men’s alcohol use (chapter 4).

Measuring social factors that are not directly observed is a key challenge in the estimation of contextual effects. I consider methods for estimating the effect of latent group factors on health outcomes using observed item responses within individuals. I compare approaches where the group-level exposure is calculated as the overall mean of items within persons within groups (aggregation and regression) with latent variable methods, namely item response modeling within individuals and structural equation modeling within groups. In particular, I explore the creation of multiple plausible values for individual perception and the use of these measures in a multilevel structural equation model. Simulation studies across a range of conditions to assess the robustness of these methods suggest that latent variable models reduce bias in the estimation of contextual effects relative to consistent attenuation in approaches based on aggregation and regression. This bias correction incurs additional variability, however. The causal model linking the latent construct to item responses affects appropriate analytic choice, as one setting in which aggregation and regression approaches perform well is when the group latent trait is a composite of individual values (formative indicators) and the sampling fraction is high. Finally, consideration of the role of third variables in affecting measurement, exposure, and outcome suggests that aggregation and regression approaches can be highly prone to bias in these scenarios, with some bias correction possible when adjusting for the aggregate value of a variable that distorts measurement. Latent variable methods provide reasonable bias correction in these situations without control for the aggregate of the third variable; adjustment for the distortion factor in the measurement model did not contribute substantially to this correction. As a whole, this work suggests that latent variable methods are worthy of further consideration in social epidemiologic analysis, with additional work warranted on reducing the variability of such methods and comparing them to a single-stage, fully latent model.

Chapter 3 presents a novel analysis of social and environmental factors related to alcohol outcomes in rural South Africa. I assess the relationship of collective efficacy and alcohol outlet density with heavy drinking and potential problem drinking using a population-representative survey of 581 young men from 22 villages in Agincourt, South Africa. In this cross-sectional analysis, informal social control and social cohesion show protective associations with men’s heavy drinking but not problem drinking: a one standard deviation difference in each factor was associated with a -4% difference in expected prevalence of heavy drinking among young men. The number of formal and informal alcohol-serving establishments per square kilometer was associated with potential problem drinking but not heavy drinking. The expected difference in prevalence of potential problem drinking was 9% for a difference of one additional alcohol outlet per square kilometer. Although preliminary, these findings lay the groundwork for further investigation of contextual causes of alcohol use and suggest that such causes could be worthwhile sites for future intervention.

The final component of this work is an assessment of a randomized community mobilization intervention intended to modify gender norms in an effort to reduce HIV transmission in Agincourt. Although reducing alcohol consumption was not a primary aim, the intervention included alcohol-related content, as inequitable masculine gender norms have been linked to risky alcohol use, including drinking before sexual activity. I test the total effect of the intervention on village prevalence of heavy drinking, potential problem drinking, and alcohol use before sex based on a follow-up sample of 575 young men. I categorize intervention engagement into low, moderate, and high doses based on the proportion of young men participating in the intervention and test the association of each dose level with each outcome as an assessment of nonlinearity. A nonlinear relationship would be consistent with the mobilization element of the intervention, which was intended to generate spillover effects from engaged individuals to peers who did not participate in the intervention. The intervention was not significantly associated with village prevalence of any of the alcohol outcomes. However, high dose of village engagement was significantly positively associated with the prevalence of heavy drinking and of potential problem drinking (14% difference and 9% difference relative to moderate engagement respectively). In contrast, the village-level relationship between increasing intervention dose and pre-sex alcohol use was protective and nonlinear, with the lowest estimated prevalence at moderate engagement, but did not reach statistical significance. These results indicate a potential harmful ecological effect of high intervention engagement on heavy drinking and potential problem drinking and are compatible with no effect or a protective effect on alcohol use before sex. Subsequent analyses should assess individual-level pathways between intervention engagement and alcohol use and consider whether intervention activities such as tavern-based discussions and large soccer tournaments could have inadvertently increased alcohol consumption.

As a whole, this work advances the methodological tools available for the analysis of contextual effects and provides new evidence of the importance of community causes of alcohol use in the context of HIV risk, an area with limited prior research. Extensions of this research can help to solidify a causal relationship from community context to alcohol use and identify ways to optimize future interventions in this area.